Skip to content

Overview#

SCALE provides a version of clangd that understands nvcc-dialect CUDA, inline PTX, and provides other CUDA-specific features.

clangd is a language server. Typically, it is run behind-the-scenes by your editor to provide autocompletion, diagnostics, etc.

In general, using SCALE's clangd in your editor is the same as the process for upstream clangd, except:

  • You set the path to the clangd executable to the one that ships with SCALE (<SCALE_DIR>/llvm/bin/clangd)
  • You add a file called .clangd to the root of your project with the following contents. This will add a few flags to the command line clangd uses to explore your code:
---
CompileFlags:
  Add: [-fno-gpu-defer-diag, -D__CUDA_ARCH__=1200]
---

These two options have the following effect:

  • -fno-gpu-defer-diag: instructs the compiler to provide diagnostics for all code, including code that seems to be unused. CUDA compilation typically ignores code for the "other side".
  • Defining __CUDA_ARCH__ tends to cause more device code to be shown to clangd. For best results, you should give clangd a combination of macros that shows it the code from both sides, all the time.

If clangd fails to autodetect which CUDA dialect you are using from the compilation database, you may need to add -fcuda-nvcc-emulation to explicitly enable the NVCC dialect.

Editor-specific tutorials for setting up SCALE-clangd are available in this section of the mnaual.